acceptodds
Under review as a conference paper at ICLR 2027

iBright-GS: Immediate Bright Gaussian Reconstruction from Low-Light Views

Abstract

Training 3D Gaussian Splatting exclusively on low-light images anchors the learned appearance to dark observations, which leads to persistently dim novel-view renderings even when the geometry is fully recovered. To tackle this issue, we propose iBright-GS, a framework that learns bright-view Gaussian appearance representations solely from low-light multi-view inputs. It first stabilizes geometric optimization via raw low-light photometric constraints to preserve the intrinsic dark appearance distribution of inputs and avoid premature photometric drift that degrades geometric fidelity. A unified training-stage bright signal is then constructed by synergistically combining a curve- and LUT-based pseudo-bright target, a dark-render brightening branch, per-Gaussian local gain, global exposure lock, and multi-view consistency. After geometry converges to a stable state with negligible structural perturbation, the constructed bright signal guides the optimization of dedicated bright Gaussian appearance coefficients. The learned bright representation can be directly rasterized at inference without additional post-processing, view-specific adjustment, or reference guidance. Extensive evaluations on LOM and RealX3D benchmarks illustrate that iBright-GS achieves state-of-the-art average PSNR and SSIM performance, with substantially higher runtime efficiency than existing advanced offline low-light 3DGS baselines.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Low Light Image Enhancement, 3D Gaussian Splatting, Novel View Synthesis

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_GVf2OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes iBright-GS, which builds on the fast reconstruction pipeline of i3DGS to reconstruct 3D scenes with normal brightness using only low-light multi-view images. The authors first reconstruct the scene geometry using the original dark images. They then combine several techniques, including image brightening, curve/LUT enhancement, per-Gaussian gain adjustment, and multi-view consistency, to generate brighter training targets. These targets are used to optimize a separate set of SH0 coefficients for normal-brightness appearance, while keeping the reconstructed geometry fixed. Experiments on LOM and RealX3D show that the method achieves higher average PSNR than previous low-light 3DGS methods while requiring substantially less runtime.

Critical strengths
  1. The related work is well discussed, the figures are clear, and the writing is easy to understand.

  2. The authors evaluated their method on both the LOM and RealX3D benchmarks for low-light 3D reconstruction, achieving promising results and outperforming some recent state-of-the-art methods.

  3. The authors conducted sensitivity analyses on the bright-appearance learning schedule and hyper-parameters, along with ablation studies of individual components, allowing readers to clearly understand which aspects contribute to the method’s strong performance.

Critical weaknesses
  1. The method mainly combines i3DGS with a separate DC color coefficient and distills pseudo-labels generated by 2D image enhancement into 3D. However, distilling 2D enhancement results into 3D is not fundamentally new. The overall approach also relies on several components, including curve/LUT, a brightening branch, per-Gaussian gain, exposure lock, and multi-view loss, which makes it difficult to identify the core technical contribution. In addition, the speed advantage largely comes from the underlying i3DGS framework rather than from a novel contribution of this paper.

  2. On the LOM benchmark, RetinexGS actually achieves better SSIM and LPIPS scores than the proposed method. Looking at the results scene by scene in Table 4, the proposed method performs worse in 4 out of the 5 scenes.

  3. As a minor suggestion, Table 3 could be split into separate results for LOM and RealX3D to help readers compare the performance more intuitively, since the two datasets feature different lighting conditions.

Questions for authors
  1. If η is not used during inference, what determines the overall brightness of the test images?

  2. Could the authors include two additional baselines: “2D enhancement + i3DGS” and “i3DGS + post-processing enhancement” to better demonstrate the effectiveness of the proposed method?

Reviewer_AqhPOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper focuses on low-light novel view synthesis, which takes in low-light views, reconstructs bright scenes, and renders bright novel views. IBRIGHT-GS split the reconstruction into 2 steps: geometric optimization via raw low-light photometric constraints, bright signal guidance further optimizes Gaussian appearance coefficients. Performance on multiple benchmarks demonstrates its effectiveness.

Critical strengths
  • The paper proposes an effective staged appearance optimization strategy that could potentially be transferred to other scenarios.

  • Performing local appearance adjustment at the 3D Gaussian level, rather than relying solely on constraints for individual views, is reasonable.

Critical weaknesses
  1. This paper seems to be an extension of i3DGS: it inherits the low-light reconstruction pipeline from i3DGS, and introduces pseudo-bright targets to optimize the appearance of the reconstructed Gaussian primitives.

  2. Based on the former point, a baseline worth comparing against is i3DGS + image-based low-light enhancement for pseudo-bright view + Gaussian color optimization under bright pseudo supervision. This can help to demonstrate how much iBright-GS improves over a simple appearance optimization pipeline.

  3. Although the ablation study of removing Bright SH0 demonstrates the effectiveness of the bright branch, it does not establish the necessity of independent bright SH0 coefficients. A more compelling comparison would be: keep the bright pseudo-supervision and frozen geometry strategy, while directly optimizing the original Gaussian color coefficients, rather than introducing independent bright SH0 coefficients.

Questions for authors
  1. Explain the uniqueness of the proposed method.

  2. Compare with a vanilla i3DGS enhancement pipeline.

  3. Demonstrate the effectiveness of

Please refer to the weaknesses for details.

Reviewer_RhRAOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2(−1)
Soundness
2
Read review
Paper summary

This paper extends i3DGS to reconstruct bright Gaussian appearance from low-light multi-view images. It first obtains a dark reconstruction with structure-gap-guided densification, then constructs a bright training signal from curve/LUT references and dark renderings, using local gains, exposure locking, and multi-view constraints. Separate bright SH0 coefficients are optimized on frozen geometry, while opacity and higher-order SH coefficients remain shared with the dark branch. Experiments on LOM and RealX3D report strong PSNR and short reconstruction times. The resulting representation can be rendered directly without additional test-time enhancement.

Critical strengths
  1. The paper addresses a practical problem of reconstructing normal-brightness views using only low-light images,without using normal-exposure images for scene optimization. The reported PSNR results are strong on both LOM and RealX3D. The improvements are also consistent across the individual scenes.
  2. The reconstruction time is considerably shorter than several offline low-light 3DGS methods. The direct rendering of bright Gaussians without additional post-processing is also useful.
Critical weaknesses
  1. The experiments do not isolate the added value of the proposed bright-signal construction. While Table 2 and Table 6 are useful, neither establishes the improvement over directly fitting the same bright SH0 coefficients to the curve/LUT pseudo-target on the same frozen reconstruction. This alternative would retain the same representation and direct inference path while removing the additional teacher-construction machinery. Removing bright SH0 entirely is not an adequate substitute, since it removes bright-appearance learning rather than comparing ways of supervising it. Without this controlled comparison, the results do not establish how much the proposed supervision contributes beyond the reconstruction backbone and direct pseudo-target fitting.

  2. The central bright-appearance learning procedure is not specified well enough to reproduce and assess. I could not find sufficient specifications of the curve/LUT predictor and the brightening operator B, or explicit definitions of Lspa and Lhist. There is also an unresolved calibration issue: Eq (17)–(18) supervise a mean-normalized rendering, whereas inference uses the unnormalized output. In image space, ignoring epsilon, a rendering proportional to the target has zero normalized error even when its raw brightness differs. Exposure locking calibrates the teacher but does not by itself explain the final output scale. Initialization and the fixed higher-order SH coefficients may provide that explanation, but their role needs clarification. These gaps concern the mechanism for obtaining directly renderable bright Gaussians, rather than peripheral implementation details.

Questions for authors
  1. How does the proposed method compare with directly optimizing bright SH0 using the curve/LUT pseudo-target on the same frozen geometry, with matched initialization and optimization budget?
  2. Since Eq (17)–(18) normalize the rendering during training but not at inference, what determines the absolute brightness of the final output?
  3. Could you provide the exact definitions and training details of the curve/LUT predictor, the brightening operator B, Lspa, and Lhist?
  4. Could you clarify how the proposed method differs technically from Wang et al.'s LLGS (ICRA 2025, different from Gan et al) and LL-Gaussian (ACM MM 2025)?
  5. Could you provide per-keyframe latency, intermediate bright-rendering results, and unordered-input evaluations to substantiate the immediate reconstruction claim?
2 comments

Thank you for the detailed rebuttal and additional experiments. The new w/o Full Teacher baseline provides stronger evidence for the benefit of the proposed bright-signal construction. I also appreciate the additional implementation details and related-work discussion. The first major concern is largely resolved and I am getting more positive about this work.

My main remaining concern is the novelty relative to i3DGS and brightness calibration. Since the online and unordered reconstruction capabilities are largely inherited from i3DGS, I think it is important to clarify whether bright Gaussian renderings are available during incremental reconstruction or only after the final finetuning. The additional timing results are helpful, but do not fully clarify this distinction. I also have a remaining question about brightness calibration. The authors' explanation is helpful, but since Eq. (17) normalizes the rendering during training but not at inference, it is still unclear to me how the absolute brightness scale of the unnormalized output is maintained.

Thank you for the additional experiments and clarifications. However, I still have concerns about the inference procedure underlying the reported results.

  • Following your clarification that η was used during evaluation, could you provide the exact inference equation used for Tables 1 and 2? The revised PDF available to me still states that η is not applied at test time, which seems inconsistent with your latest response.

  • How exactly is η computed for a held-out test view? Is it estimated separately for each view or shared across the scene? In particular, does its computation require the low-light image at the held-out viewpoint, or can it be determined solely from the training observations and learned representation? For the w/o Full Teacher baseline, which signal is used to determine the target mean?

  • Could you report PSNR, SSIM, and LPIPS with and without evaluation-time calibration, using the same learned parameters and test split? Please also clarify whether the same calibration procedure and test-time information were used for Ours and w/o Full Teacher.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.